cooper: task colocation with cooperative gamesqiuyun llull, songchun fan, seyed majid zahedi,...

Cooper: Task Colocation with Cooperative Games

Qiuyun Llull, Songchun Fan, Seyed Majid Zahedi, Benjamin C. LeePresented by: Qiuyun Llull

Duke University

HPCA – Feb 7, 2017

1

Task Colocation in Datacenters

2

Datacenters colocate applications to increase server utilization

3

Colocation Contention

Colocation interference can lead to performance degradation

4

System Setting

• Alvin, Ben, and Dan are working towards HPCA papers.• They share a cluster and divide processors equally.• Ben’s applications are memory intensive.• Alvin and Dan’s applications are not memory intensive.

5

System Setting

6

Strategic Behavior

• Alvin, Ben, and Dan are strategic.• Can smaller, separate clusters improve performance?• Alvin and Dan share separate cluster to improve performance.

7

Strategic Behavior

8

Without incentives, strategic users may…• Bypass common management policy• Migrate tasks for better colocations• Procure private machines

Strategic action fragments cluster and harms efficiency

Strategic Behavior

9

Pursues Performance• Predicts contention quickly and accurately• Colocates tasks for system performance• Colocates tasks with complementary demands

Neglects Incentives• Overlooks strategic behavior• Fails to encourage users to colocate

Prior Research

10

Stability• No group of users break away to form separate system

Satisfied Preferences• More users colocate with preferred tasks

Fair Attribution of Costs• Users that contribute more to contention suffer higher losses

Incentivizing Colocation

11

Example

PreferencesA : B > C > DB : A > C > DC : A > B > DD : C > A > B

12

Example

PreferencesA : B > C > DB : A > C > DC : A > B > DD : C > A > B

13

A framework that incentivizes strategic users to colocateby providing desirable system outcomes:

• Stability• Satisfied Preferences• Fair Attribution of Costs

Cooper

14

• System Setting• Incentivizing Colocation• Cooper Colocation Framework• Evaluation

Agenda

15

• Strategic agents are users and tasks• Utility is task performance• Colocation preferences describe preferred co-runners• If u(A,B) > u(A,C), then A prefers B over C

• Actions are -- participate or break away

Cooperative Game

16

Colocations are stable when no group of users can improve their performance by changing colocation.

Game Equilibrium

Cooper Framework

17

Query Interface

Preference Predictor

ActionRecommender

Users

Agents Coordinator

System Profiler

ColocationPolicies

Job Dispatcher

Machines

Colocation Policies

18

Query Interface

Preference Predictor

ActionRecommender

Users

Agents Coordinator

System Profiler

ColocationPolicies

Job Dispatcher

Machines

Matching people in life

19

20

Stable MatchingAlgorithm partitions tasks into two sets• Tasks in one set propose.• Tasks in other set accepts, rejects.

Task updates co-runners • Accept proposal if performance improves

Algorithm terminates when all tasks matched

[1] D. Gale and L. Shapley, “College admissions and the stability of marriage,” American Mathematical Monthly, 1962. [2] R.W.Irving, “An efficient algorithm for the stable roommates problem,” Journal of Algorithms, pp. 577–595, 1985.

21

A

F

E

D

B

C

D>F>E

E>F>D

E>D>F

A>B>C

A>C>B

C>B>A

Stable Matching

22

Stable PoliciesStable Marriage Random (SMR)• Partition tasks randomly

Stable Marriage Partition (SMP)• Partition tasks with domain-specific knowledge• Memory-intensive tasks propose

Stable Roommate (SR)• No partition• Any task proposes to any other.

23

Baseline PoliciesGreedy (GR)• Colocate tasks to minimize performance loss

Complementary (CO)• Colocate tasks with complementary resource demands

Preference Predictor

24

Query Interface

Preference Predictor

ActionRecommender

Users

Agents Coordinator

System Profiler

ColocationPolicies

Job Dispatcher

Machines

25

• Profile colocation performance with sparse samples

• Rate co-runners with profiles

• Predict ratings with collaborative filtering• Infer ratings based on task similarity• Suppose A: B > C and A is similar to D• Then D: B > C

• Construct preference list per task based on ratings

Preference Predictor

Action Recommender

26

Query Interface

Preference Predictor

ActionRecommender

Users

Agents Coordinator

System Profiler

ColocationPolicies

Job Dispatcher

Machines

27

• Assess assigned matches for each task

• Search preference list for better co-runners• Suppose X: A > B, and X matched to B• X messages A to suggest new match

• Recommend break away • Suppose A also prefers X over assigned match. • X, A should break away

Action Recommender

Cooper Recap

28

Query Interface

Preference Predictor

ActionRecommender

Users

Agents Coordinator

System Profiler

ColocationPolicies

Job Dispatcher

Machines

29

• System Setting• Incentivizing Colocation• Cooper Colocation Framework• Evaluation

Agenda

Workloads• PARSEC for multithreaded benchmarks• Spark for task-parallel machine learning

System Measurements• 10 nodes, each with 2 processors and 24 cores• Two tasks share a processor each with half the cores

System Simulation• 500 nodes with varied task populations• Simulate colocations with system profiles

30

Experimental Methods

Fair Attribution of Costs

31

Stable Marriage Random (SMR)

swap

t.bo

dytr.

dedu

pca

neal

svm

linea

rstr

eamc.

decis

iongr

adien

tna

iveco

rrelat

.0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

Thro

ughp

ut P

enal

ty

Complementary (CO)sw

apt.

body

tr.de

dup

cane

alsv

mlin

ear

strea

mc.de

cision

grad

ient

naive

corre

lat.0.00

0.05

0.10

0.15

0.20

0.25

Thro

ughp

ut P

enal

ty

Tasks that contribute more to contention suffer higher penalties

x-axis sorts applications by memory intensity

32

Satisfied Preferences

More users colocate with preferred tasks.

SR/CO SMR/CO SMP/CO

020

060

010

00N

umbe

r of A

gent

s

Improved PerformanceUnchanged PerformanceDegraded Performance

33

Stability

Fewer users break away to form separate system

0 1 2 3 4 5

020

040

060

080

0

CO

0 1 2 3 4 5

020

040

060

080

0

GR

0 1 2 3 4 5

020

040

060

080

0

SMR#

user

s br

eakin

g aw

ay

x-axis: gains (%) for which tasks break away

Loss

Rel

ative

to S

tand

-Alo

nestable matches

0.1

0.0

0.1

0.2

0.3

0.4

0.5

34

Performance

Stable colocations preserve system performance

More in the paper …Cooper Implementation• Profiler and preference predictor• Adapted matching algorithms• Action recommender and job dispatcher

Cooperative Game Theory• Shapely value for fair division• Extending beyond pairs

Experimental Results• Sensitivity to system scale and job mix• Comprehensive policy comparisons

35

ConclusionCooperative Games for Shared Systems• Formalize interactions between strategic users• Incentivize user participation• Enable fair task colocation

Management Desiderata• Fair attribution of costs• Satisfied preferences• Stability

Fairness versus Performance• Stable colocations satisfy more users• Stable colocations preserve system performance

36

Q & A

Thank you!

35